OCR Interpretability

Explanation

OCR interpretability checks, for document-extraction systems, whether each extracted field is backed by a visual overlay on the source document image - a highlight or bounding box showing exactly where on the scanned page the value was read from. It matters because OCR can misread digits or pull text from the wrong region; an overlay lets a human reviewer instantly confirm that, say, the account number on a scanned cheque or the figure on a loan application was taken from the right spot. It is measured as OCR Overlay Completeness = Fields_with_Overlay / Total_Fields, the proportion of extracted fields that have a corresponding visual overlay. To implement it, have the OCR/extraction stage record pixel coordinates (bounding boxes) for every field it outputs and render those overlays in the review UI, then compute the ratio of fields-with-overlay to total fields per document and log it as evidence. The proposed thresholds are: target ≥ 0.90, investigate when below 0.85, and remediate when below 0.80. Falling under 0.85 triggers investigation into why overlays are missing (e.g. low-confidence regions or layout failures), and dropping below 0.80 triggers remediation of the OCR pipeline before its outputs are used in downstream decisions. Aligns with RBI Understandable by Design, NIST MEASURE 2.8, and ISO Clause 7.4 (all proposed).